[Numpy-discussion] Problems building numpy

[numpy-discussion at robince.ftml.net]

Hi,

I am keen to evaluate numpy as an alternative to MATLAB for my PhD work
and possible wider use within the department. To make a fairer
comparison I wanted to build it with optimised ATLAS/LAPACK etc. hence
building from source.

I am running into some problems though.

I am using Windows XP SP2, with latest Cygwin and I'm trying to follow
the instructions on the wiki.

Firstly, is what I'm trying possible? On the Installing Scipy/Windows
page it says MinGW gcc and g77 are best supported, but also says to
build against binary Python download you need to use MSVC. From
http://boodebr.org/main/python/build-windows-extensions it seems
building extensions with gcc is fine (and I built PyCrypto successfully
as a test). So can I do what I am trying to do (build numpy/scipy on
windows using cygwin without MSVC installed) against the downloaded
Python distribution? If not, I can't find any resources about building
Python from source on Windows using Cygwin, so it seems like I would be
completely stuck.

The next problem is that although I filled in the site.cfg file as
documented (below), the setup.py script doesn't seem to pick it up and
doesn't look in any of the specified directories. I can get around this
by putting ATLAS/LAPACK libs in C:\, but obviously this isn't very
satisfactory. Also is the entry for fftw correct? I couldn't find any
information about this on the wiki.

------
[blas_opt]
libraries = f77blas, cblas, atlas
library_dirs = C:\cygwin\home\mqbxfri2\ATLAS\atlas_win32\lib
include_dirs = C:\cygwin\home\mqbxfri2\ATLAS\atlas_win32\include
#
[lapack_opt]
libraries = lapack, f77blas, cblas, atlas
library_dirs = C:\cygwin\home\mqbxfri2\ATLAS\atlas_win32\lib
include_dirs = C:\cygwin\home\mqbxfri2\ATLAS\atlas_win32\include
[fftw]
libraries = fftw3-3
library_dirs = "C:\fftw"
-----

Following this there seem to be some more problems with the setup.

"python setup.py build --compiler=mingw32" fails with:
  File
  "C:\cygwin\home\mqbxfri2\numpy_trunk\numpy\distutils\fcompiler\__init__.py",
  line 731, in _find_existing_fcompiler
    c.customize(dist)
  AttributeError: 'NoneType' object has no attribute 'customize'

c is the result of the new_fcompiler function.

"python setup.py config_fc --help-fcompiler" failes with:
  File "c:\Python25\lib\distutils\msvccompiler.py", line 270, in
  initialize
    "version of the compiler, but it isn't installed." % self.__product) 
  distutils.errors.DistutilsPlatformError: Python was built with Visual
  Studio version 7.1, and extensions need to be built with the same
  version of the compiler, but it isn't installed.

Although as I mentioned I can successfully build extensions with gcc.

Reading through the help I saw that there is a "none" fcompiler type, so
using that I get a bit further:

"python setup.py build --compiler=mingw32 --fcompiler=none" run
initially with build directory removed gives the same NoneType error,
but then running it again the build appears to start. I then run into
bug #220 http://projects.scipy.org/scipy/numpy/ticket/220:

numpy\core\src\multiarraymodule.c: In function `initmultiarray':
numpy\core\src\multiarraymodule.c:7563: error: `NPY_ALLOW_THREADS'
undeclared (first use in this function)
numpy\core\src\multiarraymodule.c:7563: error: (Each undeclared
identifier is reported only once
numpy\core\src\multiarraymodule.c:7563: error: for each function it
appears in.)

I would really appreciate any help to get this working. I understand
building numpy doesn't require a fortran compiler, but scipy does. I am
hoping to build scipy as well, so presumably the config system needs to
recognise the cygwin g77 compiler for that to work? During the config I
also see the following message:
don't know how to compile Fortran code on platform 'nt' with 'gnu'
compiler. Supported compilers are: absoft
Does this mean it isn't possible to build scipy on Windows with Cygwin
compilers?

If I am eventually successful I would be happy to update the wiki with
some more detailed instructions based on my experiences. 

Finally, I can't find any discussion of the relevant merits of ATLAS vs
MKL, other than the different licensing. Is it expected that MKL
performs better? Which is recommended?

Thanks very much,

Robin